Project
Deep Restricted Kernel Machines: Methods and Foundations
This research proposal entitled "Deep Restricted Kernel Machines: Methods and Foundations" is related to two main directions in the field of machine learning:
- deep learning
- support vector machines and kernel methods
This project aims at an in-depth study of the recently proposed "Deep Restricted Kernel Machines" (Deep RKM). A method of conjugate feature duality is used to obtain a representation in terms of visible and hidden units. In this way the class of restricted kernel machines can be linked to restricted Boltzmann machines, which do not contain hidden-to-hidden connections. Deep RKM is obtained by coupling the restricted kernel machines over different levels.
The main objectives of the proposal are
- to investigate the duality principles
- to extend the class of restricted kernel machine models
- to explore different coupling schemes and obtain efficient learning rules
- to develop methods for large scale problems and big data.
The project intends to achieve a new powerful class of machine learning techniques for supervised, unsupervised and semi-supervised learning, and contribute to setting new foundations both for deep learning and for support vector machines and kernel methods.